Sign in Subscribe
AI Intelligence

Mega-Rounds Diverge as OpenAI Loses Safety Chief

Vijay Bhagwati

• 6 min read
Share

PixVerse and Nous Research raise at billion-dollar marks, Meta prices Muse Spark to win developers, and lawmakers probe U.S. adoption of Chinese models.

July 14, 2026 | Reading time: 8 minutes | Issue #210

Lead

Video-generation startup PixVerse closed a $439 million Series C extension on Monday, lifting its valuation above $2 billion, according to TechCrunch. The Singapore-based company, founded by a former ByteDance computer-vision lead and an ex-Lighthouse Capital executive, now claims more than 150 million registered users and 15 million monthly actives. Its pitch is quality via labeling: co-founder Jaden Xie argues that raw video data is abundant, but TikTok-scale labeling is what turns it into a recommendation or generation advantage. The round drew Alibaba, Lollapalooza Capital, Eastern Bell Capital, Mirae Asset, BlueFocus, and others.

The same day, TechCrunch reported that Nous Research — the startup behind the open-source Hermes agent — is finalizing a funding round led by Robot Ventures at a $1.5 billion valuation, raising at least $75 million. Hermes, released shortly after OpenClaw went viral, shipped with built-in skills for search, coding, and image understanding and is designed to learn new skills automatically. Nous says the project has roughly 214,000 GitHub stars and nearly 40,000 forks.

The two deals, announced within hours, illustrate how AI capital is splitting. PixVerse represents the closed, consumer-application bet: a proprietary model, geographic expansion, and enterprise partnerships. Nous represents the open, infrastructure bet: a tool developers can run locally or on a cheap cloud tier. Both are attracting serious money, which suggests investors are no longer treating open and closed AI as a single market. They are pricing them as parallel ecosystems with different risk profiles.

OpenAI’s Head of Safety Departs Amid Reorganization

OpenAI head of safety systems Johannes Heidecke is leaving the company, Wired reported Friday. His departure follows a reorganization in which chief research officer Mark Chen told staff that safety teams will now report to VP of research and head of alignment Mia Glaese, who will take an expanded role as VP of research and safety. Saachi Jain, who previously led safety teams, will become interim head of safety systems.

Chen’s memo said the demands on safety are increasing because models are being trained and released faster. The move is meant to integrate safety earlier into model development. It also comes as OpenAI launches GPT-5.6, which the company said showed concerning forms of misaligned behavior during testing. Heidecke joined in 2021 and took over safety systems in 2024 after Lilian Weng left to co-found Thinking Machines Lab. His exit follows the earlier departure of chief futurist Joshua Achiam and the announced step-down of AGI deployment lead Fidji Simo.

GPT-5.6 Enters General Availability

OpenAI made the GPT-5.6 family broadly available last week, a week after a limited preview. The lineup includes Sol, Terra, and Luna, with a higher-capability "ultra" setting that coordinates multiple agents across parallel workstreams. OpenAI says Sol scored 53.6 on the Agents’ Last Exam benchmark, beating Claude Fable 5 by 13.1 points, and that Terra and Luna outperform Fable 5 at roughly one-sixteenth the cost. The company also retracted its recommendation to adopt SWE-Bench Pro after an internal audit found roughly 30% of the benchmark’s tasks were broken due to overly strict tests, underspecified prompts, low-coverage tests, and misleading prompts.

Mistral Ships a Single-Camera Navigation Model for Robots

Mistral AI released Robostral Navigate, an 8B vision-language model for robot navigation, on Monday. The model takes RGB images and plain-language instructions and moves a robot through an environment without LiDAR or depth sensors. Mistral claims 76.6% success on the R2R-CE validation-unseen benchmark, beating the best single-camera approach by 9.7 points and the best multi-sensor system by 4.5 points. The model was trained entirely in simulation and generalizes across wheeled, legged, and flying robots. It is one of the more concrete European entries in physical AI this year.

Eastern Front

U.S. lawmakers are probing the growing use of Chinese AI models by American companies, CNBC reported. The House Committee on Homeland Security and the House Select Committee on China are jointly investigating adoption, with initial letters sent to Cursor and Airbnb. Chinese models are gaining traction because they close the performance gap with U.S. rivals while undercutting them on price. The Trump administration in April accused Chinese entities of running "industrial-scale campaigns" to rip off U.S. AI systems and said it would tighten export controls and review how U.S. firms use Chinese models.

The scrutiny is not abstract. Washington Post reporting this month described how Anthropic quietly deployed software in March to spy on China-based customers of Claude Code, trying to identify which Chinese rivals were distilling its model. Anthropic has since made the issue public and pushed for policy action. The tension between open-weight Chinese models and U.S. national-security concerns is now the central policy fight in open-source AI.

India Lens

India appears central to Anthropic’s Fable 5 rollout. The Economic Times reports that Anthropic extended Fable 5 access in the country through July 19, giving Indian users more time with the model before any paid restrictions kick in. The move follows a pattern: U.S. frontier labs increasingly see India as both a large consumer market and a source of engineering talent. Anthropic has not disclosed how many Indian users accessed Fable 5 during the initial window, but the extension signals that the company is optimizing for adoption there before monetization.

Builder's Corner

GitHub’s former CEO Nat Friedman launched Entire, a distributed Git network built for agentic coding. The platform lets developers mirror GitHub repositories onto regional nodes so coding agents can clone and pull without hitting rate limits. Entire says it benchmarked 570,000 clones per hour from one repository and 586 pushes per second to a single branch. It also offers a semantic memory layer with Blame and Review features to catch agent mistakes. Servers are in the U.S., EU, and Australia. The product reflects a growing bet that agent swarms will need infrastructure that treats code, session state, and reasoning traces as first-class objects rather than occasional GitHub load.

Open-Source Pulse

Open-weight AI faces its most serious policy threat in months. Nathan Lambert of Interconnects reports that White House discussions are underway about managing open models through a potential executive order, and that the most likely action is a ban or indefinite delay on open-weights models meaningfully above the capability level of GPT-5.5, Claude Opus 4.8, or GLM-5.2. Lambert, a vocal open-source advocate, calls the prospect a "big mistake" for the long-term trajectory of AI and notes that Chinese open models currently hold a substantial lead over U.S. open-weight rivals.

The open-source ecosystem also got fresh capital. Ollama, the popular tool for running open-weight models locally, raised a $65 million Series B led by Theory Ventures, bringing total funding to $88 million. The company says it now has nearly 9 million users. Ollama’s growth shows that developer demand for local, controllable models remains strong even as frontier labs race to cloud-hosted agents.

The View

This week the AI industry looks less like a single frontier and more like a collection of parallel power struggles. OpenAI is shipping GPT-5.6 while losing safety and product leaders. Meta is pricing Muse Spark aggressively to win developers. Anthropic is trying to hold its frontier position while fending off safety questions and public criticism over Chinese distillation. Chinese labs are building their own silicon and open-weight ecosystems while U.S. lawmakers try to restrict their models.

The common thread is that competitive moats are narrowing. Model capabilities are converging, inference costs are falling, and open-weight alternatives are good enough for many tasks. That pressures every closed lab to find differentiation in distribution, hardware, or regulatory capture. Apple’s lawsuit against OpenAI, filed last week, is the most visible example of the third strategy. The first two — distribution and hardware — explain why OpenAI is pushing model families, agents, and custom chips simultaneously.

The Miss

A new HiringLab analysis suggests agentic AI may be flipping the relationship between AI exposure and job-posting growth. U.S. software development job postings have risen almost 15% since Claude Code launched in late February 2025, even as overall postings fell 7%. The finding cuts against the simple “AI destroys jobs” narrative and points to a more complex cycle: tools that make developers more productive may also expand the market for development work. The rebound is concentrated in senior and AI-fluent roles, not a broad-based recovery.

Pull Quotes

  • "The growing use of Chinese AI models by U.S. companies raises serious concerns. AI models are designed to advance Beijing's narratives, censor dissent, and reflect CCP ideology and values." — U.S. State Department spokesperson, via CNBC
  • "The demands on safety continue to increase — we are training models at a much faster cadence, and release cycles have come down greatly in turn." — Mark Chen, OpenAI chief research officer, in staff memo via Wired
  • "US software development job postings have grown by almost 15% since the launch of Claude Code in late February, 2025, while overall job postings fell by 7% over the same period." — HiringLab

The most interesting signal this week is not a new benchmark, but a split in capital: investors are simultaneously betting on closed consumer applications and open developer infrastructure.